Time and Day-Based Peak Electricity Demand Forecasting: A Comparative Analysis of Machine Learning Models for Peak Cut and Decarbonization

This study aims to address peak power cuts and shifts in Japan by forecasting peak electricity demand using machine learning models. The focus is on a method that excludes meteorological data and utilizes only time-based features. This approach enables medium-term forecasts one month ahead, making it feasible for operational planning without reliance on weather forecasts. It also reduces computational requirements, avoiding the need for high-spec central processing units or graphical processing units. We evaluate and compare several models, including Random Forest, XGBoost, K-Nearest Neighbors, Decision Tree, and LightGBM, to determine their accuracy in providing reliable forecasts. The LightGBM model, in particular, demonstrated remarkable results, achieving forecasting accuracies of 60 in June 2024 and 74 in July 2024. This indicates that it is a highly accurate and computationally efficient model, significantly enhancing peak demand forecasting precision. This approach facilitates peak cut and peak shift strategies, which help avoid electricity shortages and achieve stable demand and supply. Furthermore, it contributes to decarbonization by eliminating unnecessary waste.